Quantative Trading Framework: USD/CAD Exchange Rate & Crude Price

Luke Talman

2023-12-06

Strategy

• Momentum-based trading strategy intended to capitalize on the correlation that exists between crude prices and the CAD/USD exchange rate over certain periods
  • Utilizes a weighted 2 and 6 day Rate of Change of USO to predict USD/CAD exchange-rate directional in the near term
  • Market Entry/Exit strategy utilizing a 30 day rolling regression on USO and USD/CAD log returns

Rationale

The foundations of the trade rely on principles of currency supply and demand, and their relationship between imports and exports between different economies

• Oil and gas extraction exports represent a significant percentage of Canada’s total exports (19.2% in 2022), with a majority of product going to the US
• Canadian energy products are generally exported to the US in exchange for USD, while Canadian producers incur significant costs in CAD
• When Oil Prices are high, the USD supply increases relative to CAD within Canada, increasing the CADs relative value
• Further, strong energy prices are often accompanied by net economic growth within Canada as producers are able to increase workforce and investment
• Following the above market fundamentals:
  • Positive momentum in crude prices should be reflected in a relative increase in CAD, vice versa
  • Due to the nature the currency inflow/outflow resulting from Oil Sales and Production (transport times, fixed-price production contracts, futures market etc.), one could expect that the currency-price effect of crude price change may not be fully realized in a single trading day
    • The trading strategy within this doc seeks to explore the possibility of this price disconnect

Research

• At its foundation, this strategy relies on a positive correlation between energy prices, and the Canadian dollars relative strength

  • As highlighted below, this relationship is not always present

• Given Canada’s relative economic diversification to some other ‘petrocurrency’ countries, factors such as policy interest rates reinvestment rate in the oil industry, and general macroeconomic trends can reduce this correlation

• Recent applicable factors include:

  • Reduction in investment confidence for Canadian oil with production cap risk, midstream issues, and high marginal costs per barrel

  • Relative high US policy interest rates

• Crude prices changes have less explanatory power in USD/CAD prices changes from 2016 onward

  • This reduction in Adj. R Squared will likely reduce trade model efficacy

Model Implementation

Data

Series Selected:

• United States Oil Fund (USO)

  • Exchanged traded security intended to capture the change in USO’s net asset value

    • The funds assets are composed of crude oil futures contracts and other oil-related contracts

  • Under performance relative to WTI spot price in recent time, in part due to negative roll yield associated with period of contango in the oil market

• CAD/USD Exchange Rate

Signals & Trades

Signals Utilized

• 2 & 6 Day Rate of Change (ROC) of daily USO close price

  • The ROC used to generate a final signal is a weighted average of the two measures, with Alpha and Beta determining the 2 & 6 day weight, respectively

  • A weighted ROC of > 0 signals a short position in USD/CAD

  • A weighted ROC of < 0 signals a long position in USD/CAD

  • Combining ROCs with two different windows intents to reduce noise, given the volatility of daily returns

• 30 Day Rolling Regression of CAD/USD daily returns on USO daily returns

  • A statically significant (at Alpha = .1) and slope coefficient < 0 generates positive signal to enter/stay in the market

  • Any other combinations of signals signals a no trade/exit from the market

  • The use of these two measures intends capture the direction and significance of the impact USO returns have on CAD/USD returns on a daily basis a given point in time, using recent historical data

• Combined Signal

  • A final signal is generated that combines the long/short directional of the ROC signal, as well as the no trade/exit signal generate via the regression

    • If the regression signal does not indicated a no trade/exit, a long/short signal is generated in based on the sign of the ROC signal

    • If the regression signal indicates a no trade/exit, a signal of zero is generated

    • Corresponding USO price data is not available for all USD/CAD trading days; on days with missing price data, a 0 (no trade/exit) signal is generated

Trades

• Trades are generated using the combined signal from the previous day

  • 0 indicated that no market position should be held; no trade will occur, unless it is to close an existing position

  • 1 indicates short position USD/CAD

  • -1 indicates a long position USD/CAD

Training Period

• A training window from 2007-01-01 to 2018-12-31 was selected

  • Price data for USO begins in 2006-05-01

  • Maximizes trade data sample size

  • Includes periods of significant energy price volatility

Optimization

Four parameters are optimized within the model:

• Number of days to consider when calculating either un-weighted ROC

  • Values between 2 and 4 days for the shorter term ROC

  • Values were considered between 5 and 10 days for the longer term ROC

• Alpha and Beta values utilized combining either ROC into a single value

  • Values between .2 and .8 at .2 increments were considered, with any final combination summing to 1

• The inclusion of a rolling regression significantly increased optimization time complexity; ideally, a broader range of Alpha and Beta’s would be considered

Risk Appetite

• Minimize max drawdown length
  • Given the varying correlation between USO and CAD/USD returns, having a strategy that has shorted anticipated drawdowns (per the training set) could provide signs earlier, should the strategy start failing
    • Filter for lower 10% percentile
• Upper 10% percentileof Omegas
  • Ensure reasonable risk-to-reward
• Select highest cumulative return of the subset

Performance

Risk Appetite-Based Selection

alpha	beta	roc1	roc2	CumReturn
0.3333333	0.6666667	2	6	0.4862476
0.2500000	0.7500000	2	6	0.4668141
0.2000000	0.8000000	2	6	0.4621479

• Optimization results suggest that during the training set, smaller values for both the short and longer term ROC are favorable

• Returns and risk both seem to benefit from non-extreme alpha and beta values, suggesting the inclusion of two Rates of Change is of benefit to the model

• Fluctuation in variance of annual returns are relatively small, with returns driving larger risk-to-return differences

• Extended out-of-market periods frequently occured based upon the rolling correlation signal

  • These extended periods of disconnects are suprising, particularly before 2016 when the static relationship weaken

• Periods of increases currency exchange rate volatility are typically accompanied by increases the USO ROC measure

  • Unsurprising, given the economic foundations of the relationship

• Cumulative returns are poor considering risk; possible explanations include:

  • In its current form, the model fails to fully capture the price-impact crude prices have on the USD/CAD exchange rate

  • Crude prices are effectively priced in to USD/CAD rates on an intra-day horizon

Testing Period

• Returns were strong though COVID volatility

  • Aligns patterns obeserved within the training set

  • Suggests that some variation of this trading model may provide utlity in higher volatility periods

• In-market percent was relatively unchanged from train period

  • Does not align with general sentiment, and statistical measures suggesting the CAD has somewaht disconnected from crude prices in recent time

    • Possibly indicative of a model failure

Limitations & Learnings

Limitations

• As it stands, the model fails to consider transaction costs

• A historical-looking non-lagged rolling regression is an imperfect measure to capture whether or not pricing dynamics that the model seeks to capitalize on are occurring

  • Simply lagging USO returns a day does not solve the issue, given the 2 and 6 day ROC window

  • Further statistical testing/modelling is required to find more optimal metric to generate a no trade/exit signal

• Ideally position size would also be a function of the strength of the relationship between USO and USD/CAD prices at any given time

Learnings

• Data frequency and timeliness significantly limit the development of trading models that rely on non-financial market data sources
• Established and understood causal relationships are, unsurprisingly, challenging to profit from
• Attempting to leverage statistical measures in an applied setting where ones understanding is tested through a metric such as P&L is good at highlighting knowledge gaps